Issue #2 · 2025-08-05

Ilia's Corner

Featured story

OpenAI goes OSS—21B-/117B-param MoE family just dropped on Hugging Face

Run GPT-class reasoning on your own hardware for the first time. After two years of closed-source dominance, OpenAI quietly released Apache-2.0 quantized weights that fit on a single A100. No API calls, no rate limits, no vendor lock-in. Whether you’re prototyping an internal copilot or building a privacy-first chatbot, you now have a fully open stack to iterate on. Get the weights and starter code at https://github.com/huggingface/transformers/releases/tag/v4.55.0

reddit_tech · 2 min read

Top stories

Public APIs: the cheat-code catalog developers secretly bookmark

Skip the backend grunt work. From weather to crypto to NASA imagery, 1400+ free and freemium endpoints are one git clone away. Spin up a weekend side-project, add a live data layer to your SaaS, or mock that missing micro-service without writing a single line of server code. Start browsing at https://github.com/public-apis/public-apis

hackernews · 1 min read

Claude Opus 4.1: 74.5 % SWE-bench without changing your prompt

Anthropic’s latest model outperforms Claude 3.5 Sonnet on real pull-request tasks—multi-file refactors, debugging across repos, and surgical code edits—without new pricing or API quirks. Drop it into your existing agent loop and watch CI pass rates climb. Details and benchmarks at https://www.anthropic.com/news/claude-opus-4-1

hackernews · 2 min read

RevokeMsgPatcher: never lose a deleted WeChat message again

A one-click C# patcher disables message recall on WeChat, QQ, and TIM while enabling multi-instance logins. Perfect for keeping audit trails or just winning that “did-you-see-I-deleted-it?” argument. Grab the latest release at https://github.com/huiyadanli/RevokeMsgPatcher

hackernews · 1 min read

GitHub’s UI is getting painfully slow—here’s what’s dragging it down

Routine actions like switching PR tabs or resizing dev-tools now pile up 5-second delays. The root causes: bloated React bundles and hidden analytics scripts. If your muscle memory is frustrated, you’re not alone. Full perf breakdown at https://yoyo-code.com/why-is-github-ui-getting-so-much-slower/

hackernews · 3 min read

Tools spotlight

Sim Studio: drag-and-drop agent workflows that ship to prod

Build, test, and deploy LLM agents visually—no YAML wrestling required. Sim Studio packages your flow into Docker images ready for Kubernetes, so you can replace expensive SaaS agents with self-hosted ones in an afternoon. Try the OSS builder at https://github.com/simstudioai/sim

Agent orchestration, workflow automation, LLM prototyping

TypeScript · 6437 stars

MCP for Beginners: write your first Model Context Protocol server in 30 min

Microsoft’s new curriculum walks you through building MCP servers and clients in C#, Java, JS/TS, and Python. Perfect for exposing internal APIs or custom data sources to Claude Desktop, Cursor, or any MCP-compatible tool. Start the tutorial at https://github.com/microsoft/mcp-for-beginners

Tool integration, context injection, AI extensibility

Multi · 6674 stars

Frigate: local AI security cameras without cloud lock-in

Turn a Raspberry Pi + Coral USB into a real-time object-detection NVR. Your footage stays on-prem, alerts hit your phone in <200 ms, and you can train custom models for anything from raccoon spotting to forklift safety. Explore the docs at https://frigate.video/

Home security, industrial monitoring, privacy-preserving analytics

Python · 40 stars

Research corner

llama.cpp just made MoE offloading trivial with two new flags

New --n-cpu-moe flag lets you pin Mixture-of-Experts weights to CPU or layer-wise mix CPU/GPU, slashing VRAM use on consumer cards. Run 70 B MoE models on a single 3090 without quantization hacks. Pull the latest at https://github.com/ggml-org/llama.cpp/pull/15077

Model Optimization · ggml contributors · 2 min read

Hashcat 6.3: crack 300+ hash types on any GPU you own

Whether you’re red-teaming or recovering a forgotten wallet password, Hashcat’s open-source, GPU-accelerated engine now covers Argon2, bcrypt, and even Apple M-series Metal. Benchmark your rig at https://github.com/hashcat/hashcat

Security · Hashcat team · 3 min read

Browse the full archive · iliareingold.com